TensorFlow使用记录 (十四）： Multi-task to MNIST + Fashion MNIST
前言
后面工作中有个较重要的 task 是将 YOLOV3 目标检测和 LanNet 车道线检测和到一个网络中训练，特别的是，这两部分数据来自于不同的数据源。这和我之前在 caffe 环境下训练检测整个车身的同时还要训练车头车尾类似，只不过环境变更到了 tensorflow，尴尬的是，这个月才真正接触 TF。因此，先拿 MNIST 和 Fashion_MNIST 这两个数据集来练练手了。
数据预处理
MNIST 和 Fashion_MNIST 这两个数据集下载下来是压缩文件格式的，为了方便后面使用，先用一下代码转一下，不要问我为啥转。。。除了官方的我就见过这种的
#!/usr/bin/env python # -*- coding: utf-8 -*- # ================================================================ # @Time : 2019/10/25 10:50 # @Author : YangTao # @Site : # @File : process.py # @IDE: PyCharm Community Edition # ================================================================ import os # MNIST MNIST = '../../MNIST_data' def convert(imgf, labelf, outf, n): f = open(imgf, "rb") o = open(outf, "w") l = open(labelf, "rb") f.read(16) l.read(8) images = [] for i in range(n): image = [ord(l.read(1))] for j in range(28*28): image.append(ord(f.read(1))) images.append(image) for image in images: o.write(",".join(str(pix) for pix in image)+"\n") f.close() o.close() l.close() convert(os.path.join(MNIST, "train-images-idx3-ubyte"), os.path.join(MNIST, "train-labels-idx1-ubyte"), os.path.join(MNIST, "mnist_train.csv"), 60000) convert(os.path.join(MNIST, "t10k-images-idx3-ubyte"), os.path.join(MNIST, "t10k-labels-idx1-ubyte"), os.path.join(MNIST, "mnist_test.csv"), 10000)
数据层
#!/usr/bin/env python # -*- coding: utf-8 -*- # ================================================================ # @Time : 2019/10/25 13:50 # @Author : YangTao # @Site : # @File : dataset.py # @IDE: PyCharm Community Edition # ================================================================ import os import numpy as np import tensorflow as tf from matplotlib import pyplot as plt F_class_names = ['T-shirt/top', 'Trouser', 'Pullover', 'Dress', 'Coat', 'Sandal', 'Shirt', 'Sneaker', 'Bag', 'Ankle boot'] class MNISTplusFashion(object): data_dirM = './MNIST' data_dirF = './F_MNIST' def __init__(self, phase, batch_size=10): self.num_classes = 10 self.train_input_size_h = 28 self.train_input_size_w = 28 self.batch_size = batch_size if phase == 'train': self.dataM = open(os.path.join(self.data_dirM, 'mnist_train.csv'), 'r').read().split('\n')[:-1] self.flagM = np.zeros(shape=(len(self.dataM)), dtype=np.int) self.dataF = open(os.path.join(self.data_dirF, 'fashion_mnist_train.csv'), 'r').read().split('\n')[:-1] self.flagF = np.ones(shape=(len(self.dataF)), dtype=np.int) elif phase == 'val': self.dataM = open(os.path.join(self.data_dirM, 'mnist_test.csv'), 'r').read().split('\n')[:-1] self.flagM = np.zeros(shape=(len(self.dataM)), dtype=np.int) self.dataF = open(os.path.join(self.data_dirF, 'fashion_mnist_test.csv'), 'r').read().split('\n')[:-1] self.flagF = np.ones(shape=(len(self.dataF)), dtype=np.int) self.dataM = [d.split(',') for d in self.dataM] self.dataF = [d.split(',') for d in self.dataF] data = self.dataM + self.dataF flag = np.concatenate([self.flagM ,self.flagF],axis=0) self.num_samples = len(flag) # dataset size self.num_batchs = int(np.ceil(self.num_samples / self.batch_size)) # 向上取整 self.batch_count = 0 # batch index # np.random.seed(1) random_idx = np.random.permutation(self.num_samples) self.data = [] for index in random_idx: self.data.append(data[index] + [flag[index]]) def __iter__(self): return self def __next__(self): with tf.device('/cpu:0'): batch_image = np.zeros((self.batch_size, self.train_input_size_h, self.train_input_size_w, 1)) batch_label = np.zeros((self.batch_size, self.num_classes)) batch_tag = np.zeros((self.batch_size, 1)) num = 0 # sample in one batch's index if self.batch_count < self.num_batchs: while num < self.batch_size: index = self.batch_count * self.batch_size + num if index >= self.num_samples: # 从头开始 index -= self.num_samples batch_image[num, :, :, :] = np.array( self.data[index][1:-1]).reshape( self.train_input_size_h, self.train_input_size_w,1 ).astype(np.float32) / 255.0 # ====================== # smooth onehot label onehot = np.zeros(self.num_classes, dtype=np.float) onehot[int(self.data[index][0])] = 1.0 uniform_distribution = np.full(self.num_classes, 1.0 / self.num_classes) deta = 0.01 smooth_onehot = onehot * (1 - deta) + deta * uniform_distribution # ====================== batch_label[num, :] = smooth_onehot # self.data[index][0] batch_tag[num] = self.data[index][-1] num += 1 self.batch_count += 1 return batch_image, batch_label, batch_tag else: self.batch_count = 0 np.random.shuffle(self.data) raise StopIteration def __len__(self): return self.num_batchs def show_batch(img_batch): grid_image = img_batch[0,:,:,0] for idx, img in enumerate(img_batch): if idx == 0: continue grid_image = np.hstack((grid_image, img[:,:,0])) plt.imshow(grid_image) plt.title('Batch from dataloader') if __name__ == "__main__": val_data = MNISTplusFashion(phase='val', batch_size=10) for idx in range(val_data.num_batchs): batch_image, batch_label, batch_tag = val_data.__next__() print("sample %d," % idx, batch_image.shape, batch_label.shape, batch_tag.shape) plt.figure() show_batch(batch_image) plt.axis('off') plt.ioff() plt.show()
这里简单讲 MNIST 和 Fashion_MNIST 按照 1:1 concat 到一起后打乱输出了，每个 batch(iter) 返回三个变量：
""" batch_image: batch x 28 x 28 x 1 batch_label: batch x 10, smooth one-hot label batch_tag: batch x 1, 0 is mnist, 1 is fashion_mnist """
网络训练
#!/usr/bin/env python # -*- coding: utf-8 -*- # ================================================================ # @Time : 2019/10/25 15:30 # @Author : YangTao # @Site : # @File : dataset.py # @IDE: PyCharm Community Edition # ================================================================ import tensorflow as tf from tqdm import tqdm import numpy as np from matplotlib import pyplot as plt from dataset import MNISTplusFashion, show_batch print(tf.__version__) # 1. create data trainset = MNISTplusFashion(phase='train', batch_size=100) testset = MNISTplusFashion(phase='val', batch_size=20000) with tf.variable_scope('Input'): tf_x = tf.placeholder(dtype=tf.float32, shape=[None, 28, 28, 1], name='x') tf_y = tf.placeholder(dtype=tf.float32, shape=[None, 10], name='y') tf_flag = tf.placeholder(dtype=tf.float32, shape=[None, 1], name='flag') is_training = tf.placeholder(dtype=tf.bool, shape=None) global_step = tf.Variable(1.0, dtype=tf.float64, trainable=False, name='global_step') idxM = tf.where(tf.equal(tf_flag, 0))[:,0] idxF = tf.where(tf.equal(tf_flag, 1))[:,0] tf_yM = tf.gather(tf_y, idxM) tf_yF= tf.gather(tf_y, idxF) # 2. define Network with tf.variable_scope('Net'): # conv1 = tf.layers.conv2d(inputs=tf_x, filters=96, kernel_size=3, # strides=1, padding='same', activation=tf.nn.relu) # 96x28x28 # conv2 = tf.layers.conv2d(inputs=conv1, filters=96, kernel_size=3, # strides=1, padding='same', activation=tf.nn.relu) # 96x28x28 # conv3 = tf.layers.conv2d(inputs=conv2, filters=96, kernel_size=3, # strides=2, padding='same', activation=tf.nn.relu) # 96x14x14 # conv4 = tf.layers.conv2d(inputs=conv3, filters=192, kernel_size=3, # strides=1, padding='same', activation=tf.nn.relu) # 192x14x14 # conv5 = tf.layers.conv2d(inputs=conv4, filters=192, kernel_size=3, # strides=1, padding='same', activation=tf.nn.relu) # 192x14x14 # conv6 = tf.layers.conv2d(inputs=conv5, filters=192, kernel_size=3, # strides=2, padding='same', activation=tf.nn.relu) # 192x7x7 # conv7 = tf.layers.conv2d(inputs=conv6, filters=192, kernel_size=3, # strides=1, activation=tf.nn.relu) # 192x5x5 # conv8 = tf.layers.conv2d(inputs=conv7, filters=192, kernel_size=1, # strides=1, activation=tf.nn.relu) # 192x5x5 # classifier = tf.layers.conv2d(inputs=conv8, filters=10, kernel_size=1, # strides=1, activation=tf.nn.relu) # 10x5x5 # predict = tf.layers.average_pooling2d(inputs=classifier, pool_size=5, strides=1) # predict = tf.reshape(predict, [-1, 1]) # ====================== conv1 = tf.layers.conv2d(inputs=tf_x, filters=32, kernel_size=5, strides=1, padding='same', activation=tf.nn.relu) # 32x28x28 pool1 = tf.layers.max_pooling2d(inputs=conv1, pool_size=2, strides=2) # 32x14x14 conv2 = tf.layers.conv2d(pool1, 64, 3, 1, 'same', activation=tf.nn.relu) # 64x14x14 pool2 = tf.layers.max_pooling2d(conv2, 2, 2) # 64x7x7 pool2_flat = tf.reshape(pool2, [-1, 7*7*64]) pool2_flatM = tf.gather(pool2_flat, idxM) pool2_flatF = tf.gather(pool2_flat, idxF) with tf.variable_scope('MNIST'): fc1M = tf.layers.dense(pool2_flatM, 1024, tf.nn.relu) fc1M = tf.layers.dropout(fc1M, rate=0.5, training=is_training) fc2M = tf.layers.dense(fc1M, 512, tf.nn.relu) fc2M = tf.layers.dropout(fc2M, rate=0.5, training=is_training) predictM = tf.layers.dense(fc2M, 10) with tf.variable_scope('F_MNIST'): fc1F = tf.layers.dense(pool2_flatF, 1024, tf.nn.relu) fc1F = tf.layers.dropout(fc1F, rate=0.5, training=is_training) fc2F = tf.layers.dense(fc1F, 521, tf.nn.relu) fc2F = tf.layers.dropout(fc2F, rate=0.5, training=is_training) predictF = tf.layers.dense(fc2F, 10) # 3. define loss & accuracy with tf.name_scope('loss'): lossM = tf.losses.softmax_cross_entropy(onehot_labels=tf_yM, logits=predictM, label_smoothing=0.01) tf.summary.scalar('lossM', lossM) lossF = tf.losses.softmax_cross_entropy(onehot_labels=tf_yF, logits=predictF, label_smoothing=0.01) tf.summary.scalar('lossF', lossF) loss = lossM + lossF tf.summary.scalar('loss', loss) with tf.name_scope('accuracy'): # tf.metrics.accuracy() 返回 累计[上次的平均accuracy， 这次的平均accuracy] accuracyM = tf.metrics.accuracy(labels=tf.argmax(tf_yM, axis=1), predictions=tf.argmax(predictM, axis=1))[1] tf.summary.scalar('accuracyM', accuracyM) accuracyF = tf.metrics.accuracy(labels=tf.argmax(tf_yF, axis=1), predictions=tf.argmax(predictF, axis=1))[1] tf.summary.scalar('accuracyF', accuracyF) # 4. define optimizer with tf.name_scope('train'): optimizer_op = tf.train.AdamOptimizer(1e-4).minimize(loss, global_step=global_step) # 5. initialize init_op = tf.group(tf.global_variables_initializer(), tf.local_variables_initializer()) # 6.train saver = tf.train.Saver() save_path = './cnn_mnist.ckpt' # Set sess configuration # ============================== config GPU sess_config = tf.ConfigProto(allow_soft_placement=True) # sess_config.gpu_options.per_process_gpu_memory_fraction = 0.95 sess_config.gpu_options.allow_growth = True sess_config.gpu_options.allocator_type = 'BFC' # ============================== with tf.Session(config=sess_config) as sess: sess.run(init_op) # ================= merge_op = tf.summary.merge_all() train_writer = tf.summary.FileWriter('logs/train', sess.graph) test_writer = tf.summary.FileWriter('logs/test', sess.graph) # tensorboard --logdir=logs --host=127.0.0.1 # ================= for epoch in range(20): pbar = tqdm(trainset) train_epoch_loss = [] for train_data in pbar: _, ls, train_output, global_step_val = sess.run([optimizer_op, loss, merge_op, global_step], feed_dict={tf_x: train_data[0], tf_y: train_data[1], tf_flag: train_data[2], is_training: True}) train_writer.add_summary(train_output, global_step=global_step_val) pbar.set_description(("train loss:{:.4f}").format(ls)) for test_data in testset: acc_testM, acc_testF, test_ouput = sess.run([accuracyM, accuracyF, merge_op], feed_dict={tf_x: test_data[0], tf_y: test_data[1], tf_flag: test_data[2], is_training: False}) print('epoch: ', epoch, ' | test accuracyM: {:.3f}, test accuracyF: {:.3f}'.format(acc_testM, acc_testF)) sess.run(tf.local_variables_initializer()) # 不加上这句的话 accuracy 就是个累积平均值了 test_writer.add_summary(test_ouput, global_step=global_step_val) saver.save(sess, save_path)
""" train loss:0.9148: 100%|██████████| 1200/1200 [01:00<00:00, 19.70it/s] epoch: 0 | test accuracyM: 0.895, test accuracyF: 0.769 train loss:0.9703: 100%|██████████| 1200/1200 [00:57<00:00, 20.88it/s] epoch: 1 | test accuracyM: 0.974, test accuracyF: 0.858 train loss:0.6034: 100%|██████████| 1200/1200 [00:57<00:00, 20.92it/s] epoch: 2 | test accuracyM: 0.982, test accuracyF: 0.879 train loss:0.5603: 100%|██████████| 1200/1200 [01:00<00:00, 19.84it/s] epoch: 3 | test accuracyM: 0.986, test accuracyF: 0.890 train loss:0.6326: 100%|██████████| 1200/1200 [01:01<00:00, 19.61it/s] epoch: 4 | test accuracyM: 0.989, test accuracyF: 0.898 train loss:0.6328: 100%|██████████| 1200/1200 [00:57<00:00, 20.91it/s] epoch: 5 | test accuracyM: 0.990, test accuracyF: 0.906 train loss:0.4472: 100%|██████████| 1200/1200 [01:01<00:00, 19.64it/s] epoch: 6 | test accuracyM: 0.992, test accuracyF: 0.909 train loss:0.7271: 100%|██████████| 1200/1200 [01:01<00:00, 19.58it/s] epoch: 7 | test accuracyM: 0.993, test accuracyF: 0.914 train loss:0.4884: 100%|██████████| 1200/1200 [00:57<00:00, 20.95it/s] epoch: 8 | test accuracyM: 0.994, test accuracyF: 0.918 train loss:0.6024: 100%|██████████| 1200/1200 [01:01<00:00, 19.59it/s] epoch: 9 | test accuracyM: 0.995, test accuracyF: 0.923 train loss:0.4632: 100%|██████████| 1200/1200 [01:02<00:00, 19.24it/s] epoch: 10 | test accuracyM: 0.995, test accuracyF: 0.927 train loss:0.5326: 100%|██████████| 1200/1200 [01:03<00:00, 19.00it/s] epoch: 11 | test accuracyM: 0.996, test accuracyF: 0.930 train loss:0.5155: 100%|██████████| 1200/1200 [01:02<00:00, 20.04it/s] epoch: 12 | test accuracyM: 0.997, test accuracyF: 0.934 train loss:0.4652: 100%|██████████| 1200/1200 [01:00<00:00, 19.77it/s] epoch: 13 | test accuracyM: 0.997, test accuracyF: 0.936 train loss:0.5368: 100%|██████████| 1200/1200 [00:58<00:00, 20.63it/s] epoch: 14 | test accuracyM: 0.997, test accuracyF: 0.942 train loss:0.3480: 100%|██████████| 1200/1200 [01:01<00:00, 19.38it/s] epoch: 15 | test accuracyM: 0.998, test accuracyF: 0.942 train loss:0.3544: 100%|██████████| 1200/1200 [01:09<00:00, 17.75it/s] epoch: 16 | test accuracyM: 0.998, test accuracyF: 0.947 train loss:0.4082: 100%|██████████| 1200/1200 [01:02<00:00, 21.23it/s] epoch: 17 | test accuracyM: 0.998, test accuracyF: 0.949 train loss:0.4467: 100%|██████████| 1200/1200 [00:59<00:00, 20.02it/s] epoch: 18 | test accuracyM: 0.998, test accuracyF: 0.952 train loss:0.4393: 100%|██████████| 1200/1200 [01:00<00:00, 19.92it/s] epoch: 19 | test accuracyM: 0.999, test accuracyF: 0.955 groundtruth: [5 0 3 4 2 4 6 5 1 8] predict: [5 0 3 4 2 4 6 5 1 8] """
可以看出，Fashion_MNIST 数据集更难一点（accuracy 更低），因此可以考虑调整两个数据集的比例。
